Google Releases Gemma 4 Open-Source AI Models With Multimodal Support and Long Context Windows
Google DeepMind released the Gemma 4 family of open models today. The latest generation of the lightweight AI lineup built from the same research as the company's Gemini models adds native support for text, audio and image inputs along with context windows that reach 256,000 tokens in the larger variants.
The models ship in four sizes: E2B and E4B optimized for edge devices plus 31B and 26B A4B configurations aimed at higher-performance workloads.
According to the Gemma releases page on the Google AI for Developers site, the update rolled out on March 31 with full documentation now live. Models appeared on Hugging Face under the Google organization within the past few hours, confirming immediate availability for download and fine-tuning.
Gemma 4 emphasizes efficiency and on-device performance. The smallest E2B and E4B variants run offline with near-zero latency on smartphones, Raspberry Pi boards and Jetson Nano hardware. Google worked with the Pixel team as well as chipmakers Qualcomm Technologies and MediaTek to ensure compatibility. Larger 26B and 31B models target personal computers and servers where they handle advanced reasoning and agentic workflows that involve multiple steps of planning and tool use.
The release marks a step forward in multimodality for the open lineup. Smaller models support native audio input for speech recognition tasks while all variants process images, video, optical character recognition and chart understanding. Training covered more than 140 languages from the start. Context length scales with model size: 128,000 tokens for the edge variants and 256,000 tokens for the 26B and 31B editions.
Google described the new family on its DeepMind models page as "byte for byte, the most capable open models" and positioned them for frontier-level capabilities in compact packages. In the accompanying announcement blog post, VP of research Clement Farabet and group product manager Olivier Lacombe stated that Gemma 4 "outcompetes models 20x its size" and delivers "a new level of intelligence-per-parameter" that lets developers reach high performance with far less hardware.
The models follow the Apache 2.0 license, granting full rights to use, modify and redistribute them for any purpose including commercial applications without royalties. This represents a more permissive stance than some earlier Gemma releases. Google noted that the models undergo the same security protocols as its proprietary systems.
Since the first Gemma models launched in February 2024 the family has seen more than 400 million downloads and spawned a community ecosystem of over 100,000 fine-tuned variants. Gemma 4 builds directly on that foundation while shifting focus toward local deployment and multimodal agentic applications that can operate without constant cloud connectivity.
The variants are now live on Hugging Face and Kaggle Models for immediate use by developers.
